Information Technology Reference
In-Depth Information
Table 4.1.
Tests performed on
vote
database
Threshold
well classif.
¬
well classif.
50%
0.2
91.25%
08.33%
0.41%
0.3
90%
09.16%
0.83%
PAT
0.4
88.33%
11.66%
0.5
87.08%
12.91%
C4.5
83.75%
16.25%
OAT
91.66%
08.33%
Table 4.2.
The confusion matrix of
Vote
database using
PAT, C4.5, OAT
a b ¡- classified as
154 18 a=democrat
2
a
b
¡- classified as
a b ¡- classified as
153 19 a=democrat
1
166 6
a=democrat
66 b=republican
34
34 b=republican
67 b=republican
Table 4.3.
Tests performed on
Nursery
database
¬
well classif. 50%
Threshold well classif.
0.001
68.25%
23.80%
07.93%
PAT 0.02
67.46%
32.53%
0.03
67.46%
32.53%
C4.5
68.25%
23.80%
06.34%
OAT
72.22%
27.77%
on several thresholds: 0.2, 0.3, 0.4 and 0.5. The result of the tests is shown
in Table 4.1. The column 50% in Table 4.1 contains the percentage of objects
having a probability of 0.5 for each class value. Our results are better than the
results given by C4.5, which are presented in the same table. But our results
are equal to those given by
OAT
when the threshold is 0.2. Generally, when we
decrease the threshold, we increase the degree of dependence between attributes,
and consequently we use more attributes to construct our trees, which decreases
the number of instances on each leaf in each tree. In Table 4.1, we note that
when we decrease the threshold, our results improve and the best results are
obtained by
PAT
with a threshold of 0.2.
In Table 4.2, we show the complete confusion matrix of the Vote database
using
PAT
,C4.5and
OAT
from the left to the right, respectively. We find that
34 objects are misclassified with C4.5 when the class is
Republican
.
In Table 4.3, we present the tests performed on the
Nursery
database [20],
which has 8 discrete attributes. The class takes 5 values:
not-recom, recommend,
very-recom, priority, spec-prior
. The training set has 12960 instances, without
missing values. The test set has 126 instances where the missing values rates are
